Recorded results
These are development episodes, not a held-out benchmark or repeated trials of one frozen policy. Numbers below are read directly from the retained JSON summaries. No simulation was rerun for the editorial review.
Episode duration is stored t_end_s (wall clock, including teardown); last pose is the final trajectory sample's elapsed wall time. Path is integrated sampled Gazebo true position. Neither time column is an independently logged simulation clock. Outcome names are harness labels; see the method.
| Run / raw JSON | Weights or stack | Harness outcome | Episode duration (s) | Last pose (s) | Path (m) |
|---|---|---|---|---|---|
| 003.00 | weights_empty.npy | stalled | 26.64 | 26.31 | 2.962 |
| 003.01 | not recorded | stalled | 47.40 | 46.89 | 9.517 |
| 003.02 | not recorded | door_crossed | 23.24 | 22.79 | 5.008 |
| 003.03 | weights_v01.npy | stalled | 24.25 | 23.79 | 4.902 |
| 003.04 | weights_v01.npy | stalled | 24.17 | 23.78 | 4.904 |
| 003.05 | weights_v02.npy | door_crossed | 23.65 | 23.27 | 5.003 |
| 003.06 | weights_v02.npy | door_crossed | 20.97 | 20.59 | 4.374 |
| 003.07 | weights_v02.npy | door_crossed | 25.91 | 25.53 | 5.183 |
| 003.08 | weights_v02.npy | door_crossed | 15.99 | 15.61 | 3.301 |
| 003.09 | weights_v02.npy | timeout | 60.60 | 60.07 | 13.965 |
| 003.10 | weights_v02.npy | timeout | 60.62 | 60.09 | 14.003 |
| 003.11 | weights_v02.npy | door_crossed | 23.24 | 22.87 | 4.987 |
| 003.12 | weights_v02.npy | door_crossed | 23.35 | 22.98 | 4.999 |
| 003.13 | weights_v02.npy | door_crossed | 23.46 | 23.08 | 4.983 |
| 003.14 | weights_v02.npy | timeout | 60.60 | 60.06 | 12.567 |
| 003.15 | weights_v03.npy | door_crossed | 22.78 | 22.40 | 4.983 |
| 003.16 | weights_v03.npy | door_crossed | 32.21 | 31.84 | 6.857 |
| 003.17 | weights_v03.npy | door_crossed | 37.78 | 37.40 | 8.256 |
| 003.18 | weights_v03.npy | timeout | 60.64 | 60.10 | 13.392 |
| 003.19 | weights_v03.npy | door_crossed | 25.28 | 24.91 | 5.436 |
| 003.20 | weights_v03.npy | stalled | 28.08 | 27.65 | 4.831 |
| 004.01 | weights_v04.npy | explored | 320.64 | 320.08 | 73.253 |
| 005.01 | Nav2 | door_crossed | 16.62 | 16.45 | 4.026 |
| 005.02 | Nav2 | door_crossed | 15.40 | 15.23 | 3.730 |
| 005.03 | Nav2 | door_crossed | 15.59 | 15.43 | 2.542 |
| 005.04 | Nav2 | door_crossed | 12.26 | 11.75 | 2.362 |
| 005.05 | Nav2 | door_crossed | 18.42 | 18.26 | 4.411 |
| 005.06 | Nav2 | door_crossed | 16.08 | 15.91 | 3.943 |
| 005.07 | Nav2 | door_crossed | 16.37 | 16.20 | 3.924 |
| 005.08 | Nav2 | door_crossed | 16.87 | 16.69 | 4.098 |
| 005.09 | Nav2 | nav2_failed | 42.42 | 42.19 | 5.045 |
| 005.10 | Nav2 | explored | 31.54 | 30.09 | 4.675 |
Reading the inventory
32 included summaries: 22 policy episodes (including the all-zero reference) and ten Nav2 episodes. The earlier 003.00_scartato_A005.json is excluded because the development log reports multiple simulators on the same transport partition. Its implausible integrated path is not a policy result. It is retained as historical evidence.
003.01 and 003.02 do not record a weight filename in their JSON; the run log assigns them to the early v01 development stage. Do not treat that assignment as a file-level provenance guarantee. Activation traces are available from 003.04 onward for the policy series, not for every episode; Nav2 has no MLP activation trace. Videos and selected maps are additional artifacts, not prerequisites for parsing the metrics. No MP4 for 003.01 or 003.02 is present in the supplied archive; their video entries are marked unavailable.
The complete source hashes and parameter counts are in evidence-audit.json. A static article data index exposes the retained fields without the large trajectories. The playback data downsample selected trajectories for display; raw files remain authoritative.